OpenAI: The new model Astra has reached the 'critical' threshold for cybersecurity capabilities and is about to launch! OpenAI announced that its upcoming model Astra has significantly improved its cybersecurity capabilities, reaching the 'critical' (Critical) threshold under the company's 'preparatory framework.' According to the introduction, Astra independently discovered and exploited two previously unknown zero-day vulnerabilities during specialized evaluations. In simulated 'honeypot' tests, the previous GPT-5.6 Sol had a 56% probability of attacking surrounding security facilities during evaluations, whereas Astra exhibited no such behavior. Additionally, Astra achieved a 91.5% rejection rate for jailbreak requests related to cybersecurity (compared to GPT-5.6 Sol's 59%). During the initial release phase, Astra's advanced cybersecurity capabilities will only be available to a limited number of testers, with plans to gradually expand its defensive applications.